ml-convolution v2.0.0

convolution

Convolution using the FFT or direct algorithm.

Installation

npm install ml-convolution

Usage

One execution

import { directConvolution, fftConvolution } from 'ml-convolution';

const input = [0, 1, 2, 3];
const kernel = [-1, 1, -1];

const outputDirect = directConvolution(input, kernel); // [-1, -1, -2, 1]
const outputFFT = fftConvolution(input, kernel); // [-1, -1, -2, 1]

The functions both take an optional third argument to determine the way borders are processed. The default value, CONSTANT, will consider that the values out of the bounds are all 0. If it is set to CUT, borders will be ignored and the result will be smaller than te input by kernel.length - 1:

const outputDirect = directConvolution(input, kernel, 'CUT'); // [-1, -2]

Optimized, multiple executions

If you need to execute the convolution many times with the same kernel and input length, you should consider instead to use the class-based API:

import { DirectConvolution, FFTConvolution } from 'ml-convolution';

// const input = [0, 255, 255, 255, 255, 0, 0, 0];
const kernel = [0.1, 0.2, 0.3];

// First parameter is the size of the inputs and allows to pre-allocate an array with the correct size
const direct = new DirectConvolution(8, kernel, 'CUT');

// The convolve function mutates the same array at each execution
direct.convolve([0, 255, 255, 255, 255, 0, 0, 0]); // [ 127.5, 153, 153, 76.5, 25.5, 0 ]
direct.convolve([255, 0, 0, 255, 255, 255, 0, 0]); // [ 25.5, 76.5, 127.5, 153, 76.5, 25.5 ]

const fft = new FFTConvolution(8, kernel, 'CONSTANT');
fft.convolve([0, 255, 255, 255, 255, 0, 0, 0]); // [ 76.5, 127.5, 153, 153, 76.5, 25.5, 0, 0 ]
fft.convolve([255, 0, 0, 255, 255, 255, 0, 0]); // [ 51, 25.5, 76.5, 127.5, 153, 76.5, 25.5, 0 ]

Benchmark

With small kernels, direct convolution is usually faster:
Current results suggest that from a kernel size around 64, FFT convolution should be used.

License

MIT